p***r
发帖数: 20570 | 1 Some thoughts on one hand
This hand was played in the bbo money bridge game. You play with a robot,
your LHO also plays with a robot. The score format is total points.
You hold Ax K9x xx AKQ8xx , both white,
The first question is what to open?
Of course 1NT is a possible choice. However, this hand is probably
too strong for 1NT because of the strong suit in clubs. You don't need
a lot from partner to make 3NT. So you should open 1C, even play with
a robot.
The bidding develops :
1C p 1D 1H ?
Now what?
You have two choices, 2NT or 3C. Obviously, after 1D and 1H, this hand looks
like a NT hand. So you bid 2NT, which in robot's system shows 18-19 HCP.
now your robot partner raises you to 3NT, H2 is led, and dummy puts down:
KQxx 8x AJTxx Tx
so you play low and RHO plays HJ. Now what?
You should win this trick right away, because from the lead, you can guess
that RHO robot holds HQJxxx or HAQJxx (HJxxxx is quite unlikely)
and your human opp holds Txx or ATx (here, robot never
falsecards J when it holds JT, so it's safe to assume that the human holds
HT.
If you don't play HK now, you may not get your HK if robot holds HQJxxx.
Also, even if you duck, robot will most likely lead another low H back to
maintain the
connection, which is indeed likely because your 2NT shows a likely stopper.
Now it's natural for you to test for clubs, if it breaks 3-2, you have 11
tricks.
Under your CK, your LHO dropped C9. Should you take the C finesse?
No, this C9 can be a false card from CJ9x or a forced card from CJ9. Also,
RHO's 1H overcall make it quite unlikely that your LHO holds a singleton in
C.
So you continue with CA, Your LHO showed our and pitched a diamond. Do you
still have any chances
to make it?
Yes, you already have 3 spades, 1 H, 3 clubs and 1 D, so all you need is an
extra trick. Here from LHO's discard, he's likely to hold 4-3-5-1 shape, so
he'll be squeezed in S and D later.
Now you can't cash a club, so you play a heart back, hope gib would cash 4
more tricks and later you
can squeeze your LHO in S and D. Your LHO jumped with HT as you expected and
the robot captured it with
HQ and continued H. Does the robot make a mistake?
Yes, the robot should either return a spade or D to upset your timing and
remove your possible entry in S and break the final
squeeze.
After robot cashed 4 more tricks in H, his partner pitched two diamonds, it
exited with a low D to partner's Q and dummy's A. You play S to your SA,
cash the last club and squeeze your LHO in S and D. After some thoughts your
LHO pitched a spade and dummy's 4th spade is good. You make $3.95 from this hand.
Review, if you are confident that your RHO holds HAQJxx, you can duck HJ,
then after a possible H continuation, you can win HK, cash two clubs, found
the bad break and exit with a H, now robot can allow LHO's HT to win and LHO can smoothly shoot back a diamond. You can then play DA, try to drop DK or DQ. If it failed, you can continue DJ, try to set it up. When LHO wins Q and plays back a spade, you can win SA, cash CQ. LHO will be squeezed again. So he has to pitch another diamond. you can cross to dummy by SK, play another D, which loses to DK. Your LHO has to return a spade to your dummy's SQ and allow you to make the rest. This is probably the strongest defense, and you
still have a lot of fun to make it.
This hand shows that you should never give it up, no matter defending or
declaring. Often, you may still have a final chance to either beat it or
make it if you play well. | l****a
发帖数: 272 | | b***y
发帖数: 2804 | 3 真正最好的防守，是第四张红心都不能兑现。兑现第四张红心之后，只要庄家对牌型判
断正确，就已经铁成的了，因为庄家最终打ＣＱ的时候，左手只能留住五张牌（三张黑
桃加上方块ＫＱ），不存在方块上的猜测了。 | p***r
发帖数: 20570 | 4 I slightly revised the review part. Your analysis is actually wrong.
declarer can simply play SA, CQ and play D to DA if LHO keeps 4 spades.
【在 b***y 的大作中提到】
: 真正最好的防守，是第四张红心都不能兑现。兑现第四张红心之后，只要庄家对牌型判
: 断正确，就已经铁成的了，因为庄家最终打ＣＱ的时候，左手只能留住五张牌（三张黑
: 桃加上方块ＫＱ），不存在方块上的猜测了。
| b***y
发帖数: 2804 | 5 OK, reconstruct the play, say LHO started with 4351 shape:
Trick 1, H2 lead, RHO plays HJ, you duck
Trick 2: RHO plays low H, you win HK
Trick 3-4: you cash CAK (LHO pitches a D)
Trick 5: you play a H, RHO wins
Trick 6: RHO cashes 1 more H (LHO pitches a D), leaving 1 H un-cashed
Trick 7: RHO plays a spade, you win SA
Trick 8: you play CQ, LHO pitches 3rd D
At this point, LHO has 3 spades and 2 D, all you need to do is to play D to
DA. The "best" defense you suggested (to cash 4th H but not last one) doesn'
t work.
Now, if RHO doesn't cash 4th H, then same as above, except:
Trick 6: RHO plays a spade, you win SA
Trick 7: you play CQ, LHO pitches 2nd D
At this point, LHO has 3 spades and 3 D, you have to guess whether his
diamonds are KQx or Kxx/Qxx. That's why I said RHO shouldn't even cash 4th
heart.
OK, I suppose when LHO keeps 3 spades and 3 D, declarer can still make (on CQ dummy pitches a spade, leaving SKQ / DAJTx, then force out DKQ). But at least defender tried best to make it hard for the declarer.
【在 p***r 的大作中提到】
: I slightly revised the review part. Your analysis is actually wrong.
: declarer can simply play SA, CQ and play D to DA if LHO keeps 4 spades.
| v**********e
发帖数: 1295 | 6 挺有意思的一副牌。
以前无聊的时候经常和GIB一起打，当时还没收费，现在只有每月一号免费。
记得打过类似的结构，不过当时GIB没有帮忙兑现输墩挤同伴，这副牌或许GIB认为已经
打不宕了。另外GIB的设定好像是适合MP的，即每墩必争。
【在 p***r 的大作中提到】
: Some thoughts on one hand
: This hand was played in the bbo money bridge game. You play with a robot,
: your LHO also plays with a robot. The score format is total points.
: You hold Ax K9x xx AKQ8xx , both white,
: The first question is what to open?
: Of course 1NT is a possible choice. However, this hand is probably
: too strong for 1NT because of the strong suit in clubs. You don't need
: a lot from partner to make 3NT. So you should open 1C, even play with
: a robot.
: The bidding develops :
| p***r
发帖数: 20570 | 7 I don't see any point why you repeated the stuff I have already pointed out
in my posts. What you suggested that there might be a guess if RHO doesn't
cash his 4th H and I showed you already that there is no guess if you read
the position as 4-3-5-1.
Btw, I have no intention to compete the BM position.
to
doesn'
【在 b***y 的大作中提到】
: OK, reconstruct the play, say LHO started with 4351 shape:
: Trick 1, H2 lead, RHO plays HJ, you duck
: Trick 2: RHO plays low H, you win HK
: Trick 3-4: you cash CAK (LHO pitches a D)
: Trick 5: you play a H, RHO wins
: Trick 6: RHO cashes 1 more H (LHO pitches a D), leaving 1 H un-cashed
: Trick 7: RHO plays a spade, you win SA
: Trick 8: you play CQ, LHO pitches 3rd D
: At this point, LHO has 3 spades and 2 D, all you need to do is to play D to
: DA. The "best" defense you suggested (to cash 4th H but not last one) doesn'
| p***r
发帖数: 20570 | 8 The failure of gib is due to the wrong constraints. GIB thinks that I have
18-19 HCPs, so LHO can't hold both DKQ. Therefore, it's safe to cash all the
H tricks to avoid over tricks. The basic design of gib has many problems,
one of the key problems is that it assumes human opp always plays its system
. No any human players would always assume that opp has their bids and human
players know that opps may sometimes upgrade or downgrade their hands.
【在 v**********e 的大作中提到】
: 挺有意思的一副牌。
: 以前无聊的时候经常和GIB一起打，当时还没收费，现在只有每月一号免费。
: 记得打过类似的结构，不过当时GIB没有帮忙兑现输墩挤同伴，这副牌或许GIB认为已经
: 打不宕了。另外GIB的设定好像是适合MP的，即每墩必争。
| b***y
发帖数: 2804 | 9 Sorry, this was not my intention. I first thought that by not cashing 4th H,
RHO can give declarer a guess. Later I realized that declarer can always
make as long as he reads the distribution correctly, so I added last
paragraph to agree with you.
BM is irrelevant, no need to bring it up. It has no benefit for me, I did
what I did only because I don't want to see people battling on the board. I
much prefer the freedom of staying as an individual contributor.
【在 p***r 的大作中提到】
: I don't see any point why you repeated the stuff I have already pointed out
: in my posts. What you suggested that there might be a guess if RHO doesn't
: cash his 4th H and I showed you already that there is no guess if you read
: the position as 4-3-5-1.
: Btw, I have no intention to compete the BM position.
:
: to
: doesn'
| b***y
发帖数: 2804 | 10 我还不知道有每月一号免费这种好事。。。
我觉得robot match（3个机器人和一个真人）里的GIB很糟糕，跟传说中的相差甚远。
甚至有那种最后一墩垫错牌（而且绝对已经是明牌）的事情。做庄我也有好几次眼睁睁
看着摊牌的庄被做宕。可能用的是测试版，付了钱的GIB才是正版？
【在 v**********e 的大作中提到】
: 挺有意思的一副牌。
: 以前无聊的时候经常和GIB一起打，当时还没收费，现在只有每月一号免费。
: 记得打过类似的结构，不过当时GIB没有帮忙兑现输墩挤同伴，这副牌或许GIB认为已经
: 打不宕了。另外GIB的设定好像是适合MP的，即每墩必争。
| | | p***r
发帖数: 20570 | 11 There could be another stronger line if you take HK on the first trick. When
you cash CA and opp drops C9, you can suspect that it could be a singleton.
So you just play back H right away to cut opp's connections.
If opps cash 4 hearts, you just come back to the squeeze position.
If LHO wins HT and shoot back D, you just play DA, back to CK to check the C
position. When the bad news comes (assume LHO discards a D), you just play
another H and play LHO for 4-3-5-1. If RHO cashes all his hearts, you have
the marked squeeze. If RHO keeps one heart, the end game would be:
KQxx - JT
Ax - x Q8x
So you take SA, cash CQ, LHO is squeezed in S and D.
RHO keeps two H, the situation is rather simple to get 9 tricks.
If LHO wins HT and switches S, you can't test clubs now. So you have to play
back H right away. Now RHO still has no way to beat this contract.
Suppose RHO cashes all hearts, you still have the simple squeeze against LHO
in S and D.
Suppose RHO cashes one heart and shoot back D, you just test clubs and set
up DJ later.
Suppose RHO cash one heart and shoot back S, you still win in dummy by SK,
play C to test the C. later, LHO still has to let go 3 diamonds and gets
squeezed in D and S. So you play D to DA and set up D, LHO has to play S to
SQ to allow you win the last diamond.
Suppose RHO cashes no heart and shoot back S, you still win in dummy by SK,
play C to test the C. LHO pitches two D under your CK and CQ. Now you have
to take the double finesse in D to make it because if you play DA and DJ,
your RHO can set up his 4th spade before you set up your 4th D.
This is probably the strongest declaring and defense for this hand. The key
for declarer is to attack H before testing clubs for entry purpose. The
strongest defense is when declarer attack H, you need to attack S, which is
the key suit for declarer's communication and to set up your 4th spade.
Later, you have to play low on D to test declarer's nerve.
This hand is actually quite educational if you try to analyze it deeper and
deeper.
【在 p***r 的大作中提到】
: Some thoughts on one hand
: This hand was played in the bbo money bridge game. You play with a robot,
: your LHO also plays with a robot. The score format is total points.
: You hold Ax K9x xx AKQ8xx , both white,
: The first question is what to open?
: Of course 1NT is a possible choice. However, this hand is probably
: too strong for 1NT because of the strong suit in clubs. You don't need
: a lot from partner to make 3NT. So you should open 1C, even play with
: a robot.
: The bidding develops :
| p***r
发帖数: 20570 | 12 Still, working out all the possibilities at the table is very difficult and
I doubt any declarers in the world would think this deep and foresee all the
possibilities in one or two minutes. Still, the general principle to cut
down opps' communications is very important in hands like this one. A lot of
unmakable(unbeatable) contracts can be makable(beatable) if you read the
position well and attack opps' communications aggressively.
When
singleton.
C
play
【在 p***r 的大作中提到】
: There could be another stronger line if you take HK on the first trick. When
: you cash CA and opp drops C9, you can suspect that it could be a singleton.
: So you just play back H right away to cut opp's connections.
: If opps cash 4 hearts, you just come back to the squeeze position.
: If LHO wins HT and shoot back D, you just play DA, back to CK to check the C
: position. When the bad news comes (assume LHO discards a D), you just play
: another H and play LHO for 4-3-5-1. If RHO cashes all his hearts, you have
: the marked squeeze. If RHO keeps one heart, the end game would be:
: KQxx - JT
: Ax - x Q8x
| v**********e
发帖数: 1295 | 13 GIB的做庄的确是莫名其妙，但防守还是不错的，经常能找到唯一成功的路线。
GIB还取决于的速度模式，大概有快中慢三档吧，越慢出错的可能越小吧。
【在 b***y 的大作中提到】
: 我还不知道有每月一号免费这种好事。。。
: 我觉得robot match（3个机器人和一个真人）里的GIB很糟糕，跟传说中的相差甚远。
: 甚至有那种最后一墩垫错牌（而且绝对已经是明牌）的事情。做庄我也有好几次眼睁睁
: 看着摊牌的庄被做宕。可能用的是测试版，付了钱的GIB才是正版？
| l****a
发帖数: 272 | 14 我来问个初级问题,
上文中提到了 "double finesse", 这个词的准确意思是什么?
A, AQ10
xx
要求拿三墩, 先飞10, 再飞Q
B, AJ10x
xxx
要求拿两墩, 典型打法是, 希望LHO 有至少一张大牌
A,B 这两种是不是都是"double finesse"? 对应的中文分别是啥?
PS: 上文中提到"double finesse",那段, 虽然是 AJ10x, 但是打法应该是指A,
因为无法承受 RHO 进手兑现 H. | p***r
发帖数: 20570 | 15 The end position is
dummy: Q - AJTx
LHO: JT - KQ9
When declarer plays a low D toward dummy, if LHO plays low D, declarer has
to guess to play DJ to make it.
【在 l****a 的大作中提到】
: 我来问个初级问题,
: 上文中提到了 "double finesse", 这个词的准确意思是什么?
: A, AQ10
: xx
: 要求拿三墩, 先飞10, 再飞Q
: B, AJ10x
: xxx
: 要求拿两墩, 典型打法是, 希望LHO 有至少一张大牌
: A,B 这两种是不是都是"double finesse"? 对应的中文分别是啥?
: PS: 上文中提到"double finesse",那段, 虽然是 AJ10x, 但是打法应该是指A,
| l****a
发帖数: 272 | 16
GIB, 以前玩过, 个人感觉只能算可以玩,
后来发现这个网站, http://www.robobridge.com/
免费下载注册,本地打,结果在网上. 感觉比 GIB 要强. 欧洲人做的.
最近几次的电脑桥牌冠军, 也都是欧洲人胜出.
看那些电脑的路线, 有时候跟人类就根本不是一样的出发点.
日本人也有个软件呢, 似乎中国人没人对这个有兴趣
sorry, 楼歪了
【在 v**********e 的大作中提到】
: GIB的做庄的确是莫名其妙，但防守还是不错的，经常能找到唯一成功的路线。
: GIB还取决于的速度模式，大概有快中慢三档吧，越慢出错的可能越小吧。
| l****a
发帖数: 272 | 17
我知道你是得猜, 砸RHO单张大牌, 在这副牌是有可能的.
但是我想问的是啥叫 double finesse
(单套打法, 跟本题无关, 假设RHO 不是单张)
【在 p***r 的大作中提到】
: The end position is
: dummy: Q - AJTx
: LHO: JT - KQ9
: When declarer plays a low D toward dummy, if LHO plays low D, declarer has
: to guess to play DJ to make it.
| p***r
发帖数: 20570 | 18 Both of your examples are double finesses.
【在 l****a 的大作中提到】
:
: 我知道你是得猜, 砸RHO单张大牌, 在这副牌是有可能的.
: 但是我想问的是啥叫 double finesse
: (单套打法, 跟本题无关, 假设RHO 不是单张)
| b***y
发帖数: 2804 | 19 Both are double finesses, because two honors are missing in each case.
中文好像就叫“双飞”吧，我也不清楚，从没看过中文桥牌书，呵呵。
【在 l****a 的大作中提到】
: 我来问个初级问题,
: 上文中提到了 "double finesse", 这个词的准确意思是什么?
: A, AQ10
: xx
: 要求拿三墩, 先飞10, 再飞Q
: B, AJ10x
: xxx
: 要求拿两墩, 典型打法是, 希望LHO 有至少一张大牌
: A,B 这两种是不是都是"double finesse"? 对应的中文分别是啥?
: PS: 上文中提到"double finesse",那段, 虽然是 AJ10x, 但是打法应该是指A,
| j*******e
发帖数: 2168 | 20 Robot is just robot. N years ago I paid about $80 for Gib, how crazy.
Still, it should not be that hard to correct the wrong assumptions in the
software. After each trick, it should take the gathered info and reevaluate.
In this case, after 2 rounds of clubs, Gib knew you had a very good club
suit and thus should be able to drop the assumption that you have at least
18HCP.
Thanks for sharing, very educational hand indeed.
the
system
human
【在 p***r 的大作中提到】
: The failure of gib is due to the wrong constraints. GIB thinks that I have
: 18-19 HCPs, so LHO can't hold both DKQ. Therefore, it's safe to cash all the
: H tricks to avoid over tricks. The basic design of gib has many problems,
: one of the key problems is that it assumes human opp always plays its system
: . No any human players would always assume that opp has their bids and human
: players know that opps may sometimes upgrade or downgrade their hands.
| | | l****a
发帖数: 272 | 21 我自己把B 叫做双飞, 把A 叫做飞双, 只是不知道正式名称. | p***r
发帖数: 20570 | 22 haha, you were rich. I played against gib for a few times on okbridge about
10 years ago and found it's actually quite weak. As I said before, wbridge5
is a good program. Gib is a quite an out of date program. BBO's decision to
have incorporate gib as part of its product is a rather bad one IMO.
reevaluate.
【在 j*******e 的大作中提到】
: Robot is just robot. N years ago I paid about $80 for Gib, how crazy.
: Still, it should not be that hard to correct the wrong assumptions in the
: software. After each trick, it should take the gathered info and reevaluate.
: In this case, after 2 rounds of clubs, Gib knew you had a very good club
: suit and thus should be able to drop the assumption that you have at least
: 18HCP.
: Thanks for sharing, very educational hand indeed.
:
: the
: system
| b***y
发帖数: 2804 | 23 GIB has been computer bridge champion for many years. But I suspect the ones
they sent out for the championship match are not those that they use in BBO
robot game.
Regardless of its quality, BBO's decision to use GIB is probably based more
on business incentive than anything else.
【在 p***r 的大作中提到】
: haha, you were rich. I played against gib for a few times on okbridge about
: 10 years ago and found it's actually quite weak. As I said before, wbridge5
: is a good program. Gib is a quite an out of date program. BBO's decision to
: have incorporate gib as part of its product is a rather bad one IMO.
:
: reevaluate.
| p***r
发帖数: 20570 | 24 GIB won twice in the computer bridge championship in 98 and 99. Then
Ginsberg hasn't participated in the championship since 03. Now it's clearly
worse than jack and wbridge5 for sure. I think Ginsberg has stopped
improving gib.
I think if the author of wbridge5 opens a similar website based on wbridge5,
he may make a lot of money. Playing with wbridge5 is certainly a much more
pleasant experience.
Of course, bbo's gib is a very limited version even without any defensive
signals. Still, many bboers believe gib plays standard or UDCA based on
their very limited observations.
ones
BBO
more
【在 b***y 的大作中提到】
: GIB has been computer bridge champion for many years. But I suspect the ones
: they sent out for the championship match are not those that they use in BBO
: robot game.
: Regardless of its quality, BBO's decision to use GIB is probably based more
: on business incentive than anything else.
| p***r
发帖数: 20570 | 25 Actually if one wants to improve one's declaring skill, play some money
bridge is a good and fast way. That's actually the way for early bridge
stars to improve in their early bridge career in rubber bridge. Hamman is
probably in the last generation that is trained at rubber bridge tables.
【在 l****a 的大作中提到】
: 我自己把B 叫做双飞, 把A 叫做飞双, 只是不知道正式名称.
| b***y
发帖数: 2804 | 26 有些纯粹是bug，比方有一次防守一个红心定约，第十二墩的时候庄家出将牌，GIB手里
有一张黑桃一张梅花Ｑ，这张黑桃是第十三张黑桃，外面没人有黑桃了。现在GIB必须
垫一张牌，结果把ＣＱ垫掉，手里留这张唯一的黑桃，第十三墩庄家ＣＪ做大了。这是
神马防守？也许软件已经认为庄家做成了，从两张牌里随机地抽取一张，但是这更说明
game tree的设计有问题，不能把经过验证的事实（１００％）跟推理出来的较大可能
事件（也许９５％）区分开来。
还有，比方说对方打一门花色，明手有２和３，照说出哪张都一样，而且不影响防守路
线，但是结果你会发现根据你是出２还是３，GIB会选择续攻或是转攻。明手的牌是明
的，根本不存在false card的问题，这个明显是bug。只可惜这方面GIB的表现比较随机
，我还没有研究出来一个规律，没法加以利用。
the software. After each trick, it should take the gathered info and :
reevaluate. In this case, after 2 rounds of clubs, Gib knew you had a : very
good club suit and thus should be able to drop the assumption
【在 j*******e 的大作中提到】
: Robot is just robot. N years ago I paid about $80 for Gib, how crazy.
: Still, it should not be that hard to correct the wrong assumptions in the
: software. After each trick, it should take the gathered info and reevaluate.
: In this case, after 2 rounds of clubs, Gib knew you had a very good club
: suit and thus should be able to drop the assumption that you have at least
: 18HCP.
: Thanks for sharing, very educational hand indeed.
:
: the
: system
| b***y
发帖数: 2804 | 27 No, Hamman is not in the LAST generation. Zia is. :-)
is probably in the last generation that is trained at rubber bridge
【在 p***r 的大作中提到】
: Actually if one wants to improve one's declaring skill, play some money
: bridge is a good and fast way. That's actually the way for early bridge
: stars to improve in their early bridge career in rubber bridge. Hamman is
: probably in the last generation that is trained at rubber bridge tables.
| p***r
发帖数: 20570 | 28 gib's decision is random. It's based on a very limited sample size, usually
less than 20 hands. So it's not determined by which low card in dummy you
play. I did a lot of testings. In one case, gib raises from a small slam to
grand slam sometimes and pass the small slam other times, which is purely
random. The basic problem is that the computing sample size is too limited.
From Ginsberg's own paper, he said that gib's performance can be quite
largely improved if the sample hands size is 100 comparing with 50.
【在 b***y 的大作中提到】
: 有些纯粹是bug，比方有一次防守一个红心定约，第十二墩的时候庄家出将牌，GIB手里
: 有一张黑桃一张梅花Ｑ，这张黑桃是第十三张黑桃，外面没人有黑桃了。现在GIB必须
: 垫一张牌，结果把ＣＱ垫掉，手里留这张唯一的黑桃，第十三墩庄家ＣＪ做大了。这是
: 神马防守？也许软件已经认为庄家做成了，从两张牌里随机地抽取一张，但是这更说明
: game tree的设计有问题，不能把经过验证的事实（１００％）跟推理出来的较大可能
: 事件（也许９５％）区分开来。
: 还有，比方说对方打一门花色，明手有２和３，照说出哪张都一样，而且不影响防守路
: 线，但是结果你会发现根据你是出２还是３，GIB会选择续攻或是转攻。明手的牌是明
: 的，根本不存在false card的问题，这个明显是bug。只可惜这方面GIB的表现比较随机
: ，我还没有研究出来一个规律，没法加以利用。
| p***r
发帖数: 20570 | 29 Well, Zia is quite a singleton. There are quite a few decent rubber bridge
players who don't play much in duplicate bridge as it appears. That Fred
Chang is certainly one of them.
【在 b***y 的大作中提到】
: No, Hamman is not in the LAST generation. Zia is. :-)
:
: is probably in the last generation that is trained at rubber bridge
| b***y
发帖数: 2804 | 30 I understand. But the issues I raised had nothing to do with sample size.
They are purely logic flaws. For example, holding 13th spade and CQ, you
should NEVER pitch CQ, unless you know this is the LOWEST club remaining.
Even if the bidding indicates that partner must have CJ left, the program
should still pitch spade and keep CQ in hand, since there is difference
between inference and fact.
Same for dummy playing 2 or 3. They should just be treated as equal, and
used as same seed to trigger same samples. The sample size may be small, so
the general quality of card play may be low, but at least I want to see some
consistency, that is very important in robot match.
【在 p***r 的大作中提到】
: gib's decision is random. It's based on a very limited sample size, usually
: less than 20 hands. So it's not determined by which low card in dummy you
: play. I did a lot of testings. In one case, gib raises from a small slam to
: grand slam sometimes and pass the small slam other times, which is purely
: random. The basic problem is that the computing sample size is too limited.
: From Ginsberg's own paper, he said that gib's performance can be quite
: largely improved if the sample hands size is 100 comparing with 50.
| | | p***r
发帖数: 20570 | 31 Of course gib is quite buggy. So the right strategy is often try to declare
as many as you can. Sometimes, when gib is lucky, it can make some quite
tough hands. Overally, I don't think gib's declaring skill is comparable
with human experts. Gib's defense is actually much weaker. If one really
wants to implement an online bridge gambling website, I think he can forget
about declaring, always allow human to declare when in dummy and try to improve robert's defense and bidding as much as he can.
so
some
【在 b***y 的大作中提到】
: I understand. But the issues I raised had nothing to do with sample size.
: They are purely logic flaws. For example, holding 13th spade and CQ, you
: should NEVER pitch CQ, unless you know this is the LOWEST club remaining.
: Even if the bidding indicates that partner must have CJ left, the program
: should still pitch spade and keep CQ in hand, since there is difference
: between inference and fact.
: Same for dummy playing 2 or 3. They should just be treated as equal, and
: used as same seed to trigger same samples. The sample size may be small, so
: the general quality of card play may be low, but at least I want to see some
: consistency, that is very important in robot match.
| l****a
发帖数: 272 | 32
...
...
...
...
seed 有一定的随机性, 同样一副牌, 即使你出牌的次序不变,
因为seed 的不同, 如果 samples 比较少, 那么很可能结果就不consistent 了.
所以, 出2还是出3，不重要.
假设吧, 你手动的把这副牌打 100 次2, 100次3, 这样增大sample 的数量,
统计一下结果, 这个结果应该有更好的 consistency 的.
【在 b***y 的大作中提到】
: I understand. But the issues I raised had nothing to do with sample size.
: They are purely logic flaws. For example, holding 13th spade and CQ, you
: should NEVER pitch CQ, unless you know this is the LOWEST club remaining.
: Even if the bidding indicates that partner must have CJ left, the program
: should still pitch spade and keep CQ in hand, since there is difference
: between inference and fact.
: Same for dummy playing 2 or 3. They should just be treated as equal, and
: used as same seed to trigger same samples. The sample size may be small, so
: the general quality of card play may be low, but at least I want to see some
: consistency, that is very important in robot match.
| m****r
发帖数: 6639 | 33 我终于笑了.
【在 l****a 的大作中提到】
: 我自己把B 叫做双飞, 把A 叫做飞双, 只是不知道正式名称.
| m****r
发帖数: 6639 | 34 没有规律, 就是gib成功的地方吧.
【在 b***y 的大作中提到】
: 有些纯粹是bug，比方有一次防守一个红心定约，第十二墩的时候庄家出将牌，GIB手里
: 有一张黑桃一张梅花Ｑ，这张黑桃是第十三张黑桃，外面没人有黑桃了。现在GIB必须
: 垫一张牌，结果把ＣＱ垫掉，手里留这张唯一的黑桃，第十三墩庄家ＣＪ做大了。这是
: 神马防守？也许软件已经认为庄家做成了，从两张牌里随机地抽取一张，但是这更说明
: game tree的设计有问题，不能把经过验证的事实（１００％）跟推理出来的较大可能
: 事件（也许９５％）区分开来。
: 还有，比方说对方打一门花色，明手有２和３，照说出哪张都一样，而且不影响防守路
: 线，但是结果你会发现根据你是出２还是３，GIB会选择续攻或是转攻。明手的牌是明
: 的，根本不存在false card的问题，这个明显是bug。只可惜这方面GIB的表现比较随机
: ，我还没有研究出来一个规律，没法加以利用。
| b***y
发帖数: 2804 | 35 不是这样。Robot match的双人赛，很重要的一点是robot必须consistent/repeatable
，这样比较真人水平才公平。我研究过，前面的叫牌、出牌如果完全一样的话，robot
接下去的牌张次序都是完全固定的。所以即使有随机seed，也是由前面的叫牌出牌产生。
【在 l****a 的大作中提到】
:
: ...
: ...
: ...
: ...
: seed 有一定的随机性, 同样一副牌, 即使你出牌的次序不变,
: 因为seed 的不同, 如果 samples 比较少, 那么很可能结果就不consistent 了.
: 所以, 出2还是出3，不重要.
: 假设吧, 你手动的把这副牌打 100 次2, 100次3, 这样增大sample 的数量,
: 统计一下结果, 这个结果应该有更好的 consistency 的.
| b***y
发帖数: 2804 | 36 其实还是有规律的，比方说如果前面出牌一样，那么每次打到这一墩，GIB都是垫ＣＱ
，而不是说有时候垫黑桃有时候垫ＣＱ。所以我说是bug，非常deterministic.
【在 m****r 的大作中提到】
: 没有规律, 就是gib成功的地方吧.
| b***y
发帖数: 2804 | 37 100次2的结果完全一样，100次3的结果也完全一样，但是2和3的结果完全不同。
【在 l****a 的大作中提到】
:
: ...
: ...
: ...
: ...
: seed 有一定的随机性, 同样一副牌, 即使你出牌的次序不变,
: 因为seed 的不同, 如果 samples 比较少, 那么很可能结果就不consistent 了.
: 所以, 出2还是出3，不重要.
: 假设吧, 你手动的把这副牌打 100 次2, 100次3, 这样增大sample 的数量,
: 统计一下结果, 这个结果应该有更好的 consistency 的.
| l****a
发帖数: 272 | 38
生。
我有点晕, 如果这样的话, 不是正好说明 robot 很 consistent 吗?
虽然也许是 consistently 地错, 那是算法问题了,
不是那种 unexpected bug.
或者, 你的意思是,
如果你出2, robot 总是一致的选择打法一,
如果你出3，robot 总是一致的选择打法二 ?
如果是这样, 就有趣了.
【在 b***y 的大作中提到】
: 100次2的结果完全一样，100次3的结果也完全一样，但是2和3的结果完全不同。
| l****a
发帖数: 272 | 39
robot 厉害, 对小牌这么敏感.
同伴垫牌, 除非是 J 以上的, 我都只当作看不见.
为什么呢? 为什么呢?
【在 b***y 的大作中提到】
: 100次2的结果完全一样，100次3的结果也完全一样，但是2和3的结果完全不同。
| l****a
发帖数: 272 | 40
你不是 o/e 吧?
【在 b***y 的大作中提到】
: 100次2的结果完全一样，100次3的结果也完全一样，但是2和3的结果完全不同。
| | | o*******n
发帖数: 6500 | 41 GIB不推理的
它只是根据随即生成的小样本来找一个大概率的打法
【在 b***y 的大作中提到】
: 有些纯粹是bug，比方有一次防守一个红心定约，第十二墩的时候庄家出将牌，GIB手里
: 有一张黑桃一张梅花Ｑ，这张黑桃是第十三张黑桃，外面没人有黑桃了。现在GIB必须
: 垫一张牌，结果把ＣＱ垫掉，手里留这张唯一的黑桃，第十三墩庄家ＣＪ做大了。这是
: 神马防守？也许软件已经认为庄家做成了，从两张牌里随机地抽取一张，但是这更说明
: game tree的设计有问题，不能把经过验证的事实（１００％）跟推理出来的较大可能
: 事件（也许９５％）区分开来。
: 还有，比方说对方打一门花色，明手有２和３，照说出哪张都一样，而且不影响防守路
: 线，但是结果你会发现根据你是出２还是３，GIB会选择续攻或是转攻。明手的牌是明
: 的，根本不存在false card的问题，这个明显是bug。只可惜这方面GIB的表现比较随机
: ，我还没有研究出来一个规律，没法加以利用。
| m****r
发帖数: 6639 | 42 这也是我的理解. 但是它产生样本的时候, 难道不是根据已知的情况来产生的?
bucky出来说说, 那个CQ, 是不是其实垫了和不垫其实没有区别? 是不是gib看到了你
没有看到的?
【在 o*******n 的大作中提到】
: GIB不推理的
: 它只是根据随即生成的小样本来找一个大概率的打法
| p*********6
发帖数: 679 | 43 你说的“money bridge”就是 bbo 上的“Play Bridge For Money”吗？它跟 Robot
Tourament 区别大吗 （除了真钱外）？我打过几次0.25 robot race，感觉难赢。
【在 p***r 的大作中提到】
: Actually if one wants to improve one's declaring skill, play some money
: bridge is a good and fast way. That's actually the way for early bridge
: stars to improve in their early bridge career in rubber bridge. Hamman is
: probably in the last generation that is trained at rubber bridge tables.
| v**********e
发帖数: 1295 | 44 这个观点很同意，来美后我就一直在找可以打rubber的俱乐部，不过没有找到，这个合
法吗？还是只有在允许赌博的地区?
【在 p***r 的大作中提到】
: Actually if one wants to improve one's declaring skill, play some money
: bridge is a good and fast way. That's actually the way for early bridge
: stars to improve in their early bridge career in rubber bridge. Hamman is
: probably in the last generation that is trained at rubber bridge tables.
| p***r
发帖数: 20570 | 45 That's why I said gib is very out of dated. wbridge5 sets up a lot of good
constraints on the inferences of the previous plays, especially when in
defense.
【在 o*******n 的大作中提到】
: GIB不推理的
: 它只是根据随即生成的小样本来找一个大概率的打法
| p***r
发帖数: 20570 | 46 Yes. Huge difference. In robot tournaments, you only play against robots. In
money bridge, you play against human-robot pairs. Also, the damage of
robot tournament is limited only to your entry fee. In the money bridge game
, you may easily lose $200 to $300 in an unlucky night in 2 cents games,
which may really push you to play better.
【在 p*********6 的大作中提到】
: 你说的“money bridge”就是 bbo 上的“Play Bridge For Money”吗？它跟 Robot
: Tourament 区别大吗 （除了真钱外）？我打过几次0.25 robot race，感觉难赢。
|
|
